* fix(ai): leave room for a full answer and explain a cut-off one
max_tokens 300 cut answers off once the reason filled its 300-character
limit, and the run showed JSON.parse's position instead of the cause.
Raise it to 1024 and turn a parse failure into a readable error, naming
the token limit when the provider reports finish_reason length.
* feat(backend): add a shared logger with levels and greppable fields
One line per event with a level, a scope and key=value fields, or one
JSON object per line with LOG_FORMAT=json. LOG_LEVEL drops lines below
it. Errors log their message, status and code, plus the stack at error
level, and bearer tokens are redacted. Nothing outside it changes yet.
* feat(ai,import): log categorization and import runs through the logger
A bulk AI run kept only its last error in memory, so which models failed
and why was lost. Each provider answer it can't use now logs the model,
finish_reason, completion and reasoning tokens and the end of the answer;
each model's outcome and the run's start and totals are logged too.
Collection, Thingiverse, Printables, profile and zip imports dropped a
failed model without a trace; they now log it with its reason, plus the
job's start and totals. The links runner's console lines move to the
logger.
* fix(ai): read an answer wrapped in markdown or prose
Xiaomi MiMo ignores response_format and sent the object in bold,
`**{...}**`, which failed as not valid JSON although the answer was
complete. When the whole text doesn't parse, try the outermost braces.
Tautvydas Derzinskas